KDCube
← AI Industry News

Managed Agent Harnesses Go Mainstream: OpenAI's Agents API & a Sovereign MCP ...

OpenAI opened its Agents API broadly in public beta, turning the Codex harness into a managed product with durable sessions, subagents, and hosted or bring-your-own sandboxes. Meanwhile Google's ADK 1.0 for Kotlin pushed agents to the edge and Huawei's Agentic Cloud exposed 5,000+ MCP tools — as the st...

signals03
sources04
watch02

Highlights

  • OpenAI's Agents API is now open broadly to API developers in public beta — turning the Codex harness into a product with durable sessions, subagents, and a choice of hosted, integrated, or bring-your-own sandbox. OpenAI runs the sessions, orchestration, context compaction, and recovery; there's no separate Agents API fee, so you pay through the models and tools each session burns. (OpenAI)
  • Google's Agent Development Kit 1.0 for Kotlin reached feature parity with the Python and Java SDKs and added on-device and hybrid AI support for Android and JVM apps — pushing agent runtimes toward the edge. (AI Agent Store)
  • Huawei's Ascend stack unveiled ThinkPro (an atomic execution/state unit for agents in openEuler) and an Agentic Cloud exposing "5,000+ general-purpose MCP assets and 1,000+ industry-specific" tools across shared 10,000-NPU clusters. (AI Agent Store)
  • Automaid shipped a persistent agent operations hub that lets agents respond to external triggers (Gmail, Slack) and act across 1,000+ integrations, HTTP APIs, webhooks, and MCP servers from the cloud. (AI Agent Store)

Key Signals

  1. OpenAI makes the harness a product

    Public beta, mid-to-late September 2026

    The Agents API hands your app the Codex harness through an OpenAI-managed endpoint: durable sessions stream progress across turns, subagents fan out and merge findings, and you decide whether execution runs on your infra, an integrated sandbox provider, or OpenAI's hosted sandbox. For operators, that collapses a lot of glue code — at the cost of trusting OpenAI to manage session context and recovery. (OpenAI)

  2. Agent runtimes reach for the edge and for sovereignty

    Sept 20-21, 2026

    Google's ADK 1.0 for Kotlin brings on-device and hybrid inference to Android/JVM builders, while Huawei's Agentic Cloud stakes out a non-Western agent stack with thousands of MCP-addressable tools. Different geographies, same lesson: MCP is becoming the default wiring for tool ecosystems, and runtimes are diversifying beyond the hyperscaler cloud. (AI Agent Store)

  3. Ops surfaces keep hardening

    Sept 20-21, 2026

    Automaid's trigger-driven hub and Swarms' Auto Agent Builder (batch runs to 500 tasks, per-completion permalinks for audit tracing and cost attribution) show the market maturing from "spin up an agent" to "run and account for a fleet." (AI Agent Store)

Why It Matters / What To Watch

  1. Managed harnesses redraw the build/buy line — and the trust surface.
    • If you adopt the Agents API, decide deliberately between OpenAI's hosted sandbox and a bring-your-own execution environment; the convenience of managed orchestration is also lock-in. (OpenAI)
    • Durable, managed sessions inherit the context-compaction risks surfaced earlier this week across the field — audit what a long-running agent silently carries forward before you trust its recovered state. (OpenAI)
  2. MCP is the connective tissue now — mind the spec churn.
    • The 2026-07-28 MCP spec went stateless (no initialize handshake, header-based routing, cacheable list results), which is exactly why hosted, load-balanced agent fleets — from OpenAI's sessions to Huawei's 5,000+ MCP assets — can scale behind ordinary infrastructure. (MCP Blog)
    • Track the migration items: Client ID Metadata Documents (CIMD) replacing deprecated Dynamic Client Registration, and Roots/Sampling/Logging on a 12-month sunset. Connectors that ignore these will break within the year. (MCP Blog)

Quick Links